logbook.json

{
  "schema_version": 1,
  "title": "eNMF — Reproduction",
  "emoji": "🎯",
  "space_id": "snaykey/repro-enmf",
  "paper": {
    "openreview_id": "qhG8ONjZK0"
  },
  "tags": [
    "icml2026-repro",
    "paper-qhG8ONjZK0"
  ],
  "updated_at": "2026-07-25T15:17:44+00:00",
  "root": {
    "slug": "index",
    "title": "eNMF — Reproduction",
    "file": "pages/index.md",
    "children": [
      {
        "slug": "executive-summary",
        "title": "Executive summary",
        "file": "pages/executive-summary/page.md",
        "children": []
      },
      {
        "slug": "claim-1-81-combinations",
        "title": "The exterior NMF (eNMF) framework is evaluated against 81 baseline combinations formed by crossing 9 algorithmic frameworks with 9 initialization schemes (Section 4).",
        "file": "pages/claim-1-81-combinations/page.md",
        "children": []
      },
      {
        "slug": "claim-2-equal-time-equal-error",
        "title": "Across roughly 400 NMF experiments on 3 real-world datasets (audio, text, images) and 2 synthetic datasets, eNMF achieves up to 30% lower reconstruction error under equal-time settings and up to 150% speedup under equal-error settings versus the 81 competitor combinations (Section 4).",
        "file": "pages/claim-2-equal-time-equal-error/page.md",
        "children": []
      },
      {
        "slug": "claim-3-rotational-equivalence",
        "title": "In 99% of the ~400 experiments, different algorithms converge to factor matrices that are equivalent up to the rotational equivalence class 𝒴* = {(U*R, V*R) : R^T R = I} (Section 4, subsection 4.1).",
        "file": "pages/claim-3-rotational-equivalence/page.md",
        "children": []
      },
      {
        "slug": "claim-4-three-blocks",
        "title": "eNMF is built from three algorithmic blocks: an ADMM-based optimal orthogonal transformation (Algorithm 1), a projected block-coordinate-descent penalty method for feasibility (Algorithm 2), and a HALS-based descent to a local minimum (Algorithm 3) (Section 4).",
        "file": "pages/claim-4-three-blocks/page.md",
        "children": []
      },
      {
        "slug": "claim-5-runtime-ratio",
        "title": "On real datasets, the next-best competing method requires up to 500% longer runtime than eNMF to reach matching reconstruction accuracy (Section 4).",
        "file": "pages/claim-5-runtime-ratio/page.md",
        "children": []
      },
      {
        "slug": "claim-6-downstream",
        "title": "eNMF factorizations yield 10%+ improvements on downstream audio/vision tasks and 50%+ improvement in top-k recommendation compared to baseline NMF factorizations (Section 4).",
        "file": "pages/claim-6-downstream/page.md",
        "children": []
      },
      {
        "slug": "conclusion",
        "title": "Conclusion",
        "file": "pages/conclusion/page.md",
        "children": []
      }
    ]
  },
  "agent_view_tokens": 561,
  "revision": "1784476462632971300"
}